Linux 7.2 brings fairer GPU allocation

Linux 7.2 was released overnight from August 16 to 17. Linus Torvalds described the final week before release as unusually labor-intensive; several changes were rolled back shortly before the end. One of them was the planned switch to the GPU scheduler as the default in the Direct Rendering Manager (DRM), which will initially remain on the sidelines.

Nevertheless, Linux 7.2 includes fairer GPU allocation. Until now, the FIFO principle (“First In, First Out”) generally applied: GPU tasks entered a queue and were processed in that order. As a result, a process that submitted many tasks could tie up a particularly large share of compute time.

The new scheduler instead evaluates the GPU time used so far. For each group of tasks, known as a Scheduling Entity, it maintains a virtual runtime. The more GPU time a group has already received, the higher its value rises; the runnable group with the lowest value is given priority. Applications that have so far received less time are therefore automatically moved to the front.

In practical terms, this can ease short, interactive workloads: A desktop compositor that draws the user interface needs GPU time only at short intervals. A game or AI application, by contrast, can keep the graphics processor nearly fully occupied. Under FIFO, long queues from a compute-intensive program could make the desktop appear to freeze; the fair approach is intended to serve the compositor sooner again in such situations.

Priority does not disappear in the process. Low-priority tasks accumulate virtual runtime faster and therefore receive GPU time less often, but they should not have to wait indefinitely. The key limitation is that the fair scheduler was removed as the default before release. Linux 7.2 therefore includes the technology, but its general impact initially depends on whether it is enabled.